SOLRICKS's picture
Restore animation app on the validated Gradio runtime
eed854f verified
|
Raw History Blame Contribute Delete
2.96 kB

A newer version of the Gradio SDK is available: 6.32.0

Upgrade
metadata
title: 3D Animation Style MiniMax H3
emoji: 🎬
colorFrom: pink
colorTo: indigo
sdk: gradio
sdk_version: 6.26.0
python_version: '3.12'
app_file: app.py
pinned: false
short_description: 3D animation style video generation with MiniMax H3
suggested_hardware: zero-a10g
models:
  - MiniMaxAI/MiniMax-H3
  - SOLRICKS/3D-Animation-Style-MiniMax-H3
  - lightx2v/Minimax-h3-Turbo
tags:
  - minimax-h3
  - lora
  - 3d-animation
  - text-to-video
  - image-to-video

3D Animation Style MiniMax H3

Create stylized 3D animation from text or a starting image with SOLRICKS 3D Animation Style.

  • Trigger: 3dsrx, added automatically.
  • Recommended LoRA strength: 0.8 for T2V, 0.4 for I2V.
  • Animals, fantasy characters and environments: try 1.0.
  • Video and synchronized native audio, 24 fps.
  • 8-step Turbo acceleration using the non-768p v1.0 FL2V LoRA.
  • 4–8 seconds requested duration, aligned to H3's supported frame counts.
  • I2V center-crops a separate copy of the image to the selected canvas.

Space owner setup

Keep the existing ZeroGPU hardware allocation. Add a Secret named HF_TOKEN under Settings → Variables and secrets. Use a read token from an account with access to SOLRICKS/3D-Animation-Style-MiniMax-H3. Do not put the token in app code, README, a public variable or the repository.

The gated style adapter is checked before the large base model is downloaded. If access is unavailable, the interface remains available and explains the required setup. After adding the secret, restart the Space.

Runtime

The MiniMax H3 transformer and VAEs run in this Space. The text encoder runs in the shared Qwen3-VL conditioner, following the split deployment used by the MiniMaxAI reference Space. Prompts and optional starting images are sent to that public conditioner. Only the visitor's ZeroGPU quota header is forwarded; the owner's HF_TOKEN is used only for the gated model download and is never sent to the conditioner. The reference deployment's image-conditioned prompt guard is preserved.

Both adapters use PEFT, so changing style strength does not repeatedly round merged base weights. The loader validates target shapes and preserves alpha/rank, fused QKV ordering and the reference model's SwiGLU conversion. The diffusers revision is pinned to match the reference deployment.

Model terms

The underlying model and the style adapter are governed by the MiniMax H3 Community License Agreement and the adapter's NOTICE. No model weights are redistributed in this Space repository.

Built by SOLRICKS.